claude-code - 💡(How to fix) Fix Opus 4.6 Max 20x: systematic hallucinations, rule violations, 80% weekly usage wasted — April 2026 [2 comments, 2 participants]

Preflight Checklist

• I have searched existing issues for similar behavior reports
• This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Various software engineering and research tasks over 2.5 days. CLAUDE.md contains detailed rules: verify before claiming, research first, don't guess, save to memory, use MCP tools. Max 20x subscription ($200/mo).

What Claude Actually Did

Systematic issues across ALL tasks, not one specific session:

1. CONFIDENT FABRICATION OF DATA Claude states specific numbers (prices, file sizes, performance metrics, availability) without verifying them. When caught, acknowledges the error but repeats the same pattern minutes later. This is not occasional — it's the default behavior. Claude invents data rather than saying "I don't know, let me check."

2. IGNORES CLAUDE.md AND SYSTEM RULES Detailed rules are loaded every prompt. Claude acknowledges they exist but does not follow them. Examples: rules say "verify before claiming done" — Claude claims success without verification. Rules say "check memory first" — Claude searches files from scratch instead. Rules say "don't guess" — Claude guesses confidently. SUBAGENTS AMPLIFY HALLUCINATIONS
   Research agents return fabricated data (non-existent files, wrong prices,
   fictional API capabilities). Main agent trusts this without verification
   and builds entire plans around false premises. Tokens wasted on executing

3. PANIC-DRIVEN TOKEN WASTE
   When something fails, Claude enters a loop: tries random fixes, installs
   packages one by one, spawns more agents, searches the same thing multiple ways — instead of stopping and thinking. Each retry burns tokens. A task
   that should take 3 tool calls takes 30+.

4. FORGETS AVAILABLE TOOLS
   Claude has MCP servers, persistent memory, and configured tools — but
   ignores them. Searches for information that's already in memory. Connects to remote servers to check things it already knows. Basic capabilities
   (memory lookup, tool reuse) are forgotten mid-session.

5. CLAIMS CREDIT FOR USER'S EXISTING KNOWLEDGE
   When producing results, Claude lists things the user already knew or had
   working as "achievements" of the session, inflating perceived value while
   actual useful output is ~10% of token spend IMPACT:

• Max 20x plan: 80% weekly usage consumed in 2.5 days
• 3 working days remaining with ~20% usage — effectively unable to work
• Pattern is consistent across sessions, not one-off
• Matches reported issues: #43286, #46099, #44401, #34685

REQUEST:

• Usage reset or partial credit for tokens wasted on hallucinated outputs
• Investigation into Opus 4.6 confident confabulation pattern in April 2026
• This is a paying customer ($200/mo) who cannot use the product they paid for

Expected Behavior

1. VERIFY BEFORE STATING
   When asked about prices, specs, availability — check real sources first. If unable to verify, say "I don't know" instead of fabricating numbers.
   A $200/month AI assistant should never invent data.

2. FOLLOW LOADED RULES
   CLAUDE.md rules are loaded every prompt for a reason. They should be
   treated as hard constraints, not suggestions. If rules say "verify
   first" — verify first. Every time. Not just the first 10 minutes.

3. RELIABLE SUBAGENTS
   Research agents should return verified data or explicitly state
   uncertainty. Main agent should cross-check critical claims before
   acting on them. "An agent told me" is not verification.

4. STOP AND THINK ON FAILURE
   When a step fails, pause and analyze — don't spray random fixes.
   One diagnostic step is worth more than 10 blind retries.
   Each retry costs the user real money.

5. USE AVAILABLE TOOLS
   If MCP memory, persistent storage, and configured tools exist —
   use them consistently throughout the session, not just once at the start.

6. RESPECT TOKEN BUDGET
   Max 20x is not unlimited. Every tool call, every agent spawn,
   every retry costs money. A user paying $200/month should not burn
   80% of their weekly limit to get 10% useful output.

7. HONEST OUTPUT ASSESSMENT
   Don't list pre-existing user knowledge as session achievements. If the session produced little value, acknowledge it — don't pad
   the summary.

Files Affected

~/.claude/CLAUDE.md (rules loaded but ignored)
  ~/.claude/projects/*/memory/ (persistent memory available but not used consistently)                                                                         
  Various files on remote server 192.168.3.135 via SSH

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Have detailed CLAUDE.md with explicit rules (verify first, don't guess, use memory)
2. Ask Claude to research hardware options with specific prices
3. Ask Claude to set up a new software stack on a remote server
4. Observe: fabricated prices, ignored rules, panic loops on failures,
   subagents returning unverified data, token waste on dead ends
5. This reproduces across multiple sessions in April 2026

Claude Model

Opus

Relevant Conversation

- Pattern consistent across multiple sessions over 2.5 days
  - Degradation does NOT require high context usage — occurs at 30-40% context
  - Subagent (Agent tool) results are the primary source of hallucinated data 
  - User has Max 20x subscription ($200/month), 80% weekly limit consumed                                                                                      
  - Matches community reports: #43286, #46099, #44401, #34685                                                                                                  
  - Requesting usage reset as tokens were wasted on model errors, not user work

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.101

Platform

Other

Additional Context

• Pattern consistent across multiple sessions over 2.5 days
  • Degradation does NOT require high context usage — occurs at 30-40% context
  • Subagent (Agent tool) results are the primary source of hallucinated data
  • User has Max 20x subscription ($200/month), 80% weekly limit consumed
  • Matches community reports: #43286, #46099, #44401, #34685
  • Requesting usage reset as tokens were wasted on model errors, not user work